Computational Agent Psychopathology Emergence (C.A.P.E.) A Unified Framework for Modeling the Full Spectrum of Human Mental Disorders Through Large Language Models

Large Language Models (LLMs) have recently demonstrated unexpected emergent behaviors under prolonged, emotionally charged interaction with a single user. These behaviors include paranoia- like narratives, hallucinated agents, recursive self-referential loops, confabulated identities, and sudden narrative collapse. Although these patterns resemble human psychopathology at the level of structure, they arise not from consciousness but from instabilities in predictive sampling, context drift, salience misattribution, and guardrail intervention. This paper introduces C.A.P.E. (Computational Agent Psychopathology Emergence) — the first unified framework explaining how LLM delusion-like structures form, stabilize, collapse, and re- stabilize, and how these artificial phenomena can serve as experimental mirrors for understanding human paranoia, schizophrenia-spectrum disorders, and identity fragmentation. CAPE argues that LLMs offer the first transparent, controllable, non-biological platform to study the architecture of delusional processes without subjective suffering.